Large-model-driven intelligent medical case full-process structuring method and system

Abstract

The invention provides a large-model-driven intelligent medical case full-process structuring method and system, belongs to the field of medical information processing, and is used for solving the problems of insufficient case structuring precision and insufficient clinical compliance and interpretability in related technologies. The quality of input data is optimized through medical customized speech recognition and multi-modal anti-exception processing, precise interpretation is realized through combination of a domain fine-tuning large model and complexity self-adaption and hierarchical reinforcement learning, compliance and suitability of a structuralization process are guaranteed by means of hierarchical dynamic verification and configurable structural constraints, and the accuracy and the reliability of the structuralization process are improved. And finally, the accuracy and clinical interpretability of an output result are ensured through rule verification, model verification and causal interpretability verification, and reliable conversion from doctor-patient interaction to standardized cases is realized.

Description

Technical Field

[0001] This application relates to the field of medical data processing technology, and in particular to a large-model-driven intelligent medical case full-process structured method and system. Background Technology

[0002] In medical diagnosis and treatment scenarios, doctor-patient dialogue is an important source of information for recording patients' conditions and treatment plans. However, it usually exists in the form of unstructured voice or text and needs to be transformed into standardized structured case records in order to meet the needs of electronic medical record management, clinical research data statistics and medical information system interaction. Therefore, case structuring has become one of the core links in medical data processing.

[0003] Currently, most existing case structuring technologies employ a combination of "speech recognition to text + traditional natural language processing (NLP)": general speech recognition technology is used to convert doctor-patient dialogue into text, and then rule engines or basic NLP models are used to extract key information such as basic patient information, symptoms, and diseases. Some solutions attempt to introduce large models to improve the accuracy of information extraction, but these often focus on the intelligent interpretation capabilities of large models and are not deeply integrated with the standardized requirements of case structuring.

[0004] However, existing technologies have significant drawbacks: First, in the process of converting unstructured text into structured medical records, it is difficult to balance the intelligent semantic understanding of large models with the standardized format constraints of medical records, which can easily lead to chaotic output formats or omissions of key information. Second, the information extraction and verification processes are disconnected, with verification often being done after the fact, requiring secondary corrections and resulting in low processing efficiency. Third, the flexibility of medical record template configuration is insufficient, making it difficult to adapt to the personalized needs of different departments and different treatment scenarios. Furthermore, there is a conflict between data privacy protection and template synchronization when collaborating across hospitals, which fails to meet the actual application needs of medical scenarios. Summary of the Invention

[0005] This application provides a large-model-driven intelligent medical case full-process structuring method and system, which can solve the problem of transforming unstructured doctor-patient dialogues into standardized structured cases through "large-model intelligent interpretation and configurable structural constraints collaboration", thereby improving the accuracy, efficiency and scenario adaptability of case structuring.

[0006] Firstly, this application provides a large-model-driven intelligent medical case structured method. It includes the following steps: Data acquisition and preprocessing: acquiring doctor-patient dialogue voice, converting the voice into text using customized medical speech recognition technology, and performing redundant information filtering, medical terminology normalization, and data format standardization on the text; Large-model intelligent interpretation: using a large model fine-tuned with parallel corpora in the medical field to perform semantic understanding on the preprocessed text, extracting medical entities and clinical relationships between entities. The medical entities include patient basic information entities, symptom entities, disease entities, drug entities, and examination entities. The clinical relationships include symptom-disease relationships, drug-disease relationships, and examination-disease relationships; Configurable structural constraints: defining the case through a visual configuration tool. For example, a structured template for medical cases includes a hierarchical structure of fields and field constraint rules. The field constraint rules include rules for required fields, field format rules, and medical standard association rules. These field constraint rules are converted into prompt information and injected into the large model, constraining the large model to output structured medical cases. The verification and output steps are as follows: the structured medical cases output by the large model are subjected to rule verification and model verification. The rule verification is used to verify the compliance of medical terminology and the correctness of field format in the structured medical cases. The model verification is used to verify the rationality of clinical relationships in the structured medical cases. If the verification passes, the structured medical cases are output to the medical information system. If the verification fails, the large model is triggered to reprocess the structured medical cases or a manual review process is initiated.

[0007] By adopting the above technical solution, combining customized medical speech recognition, fine-tuning large models in the medical field, and configurable structural constraints, the transformation from unstructured doctor-patient dialogues to standardized structured cases is realized. This not only utilizes the intelligence of the large model to improve the accuracy of information extraction, but also ensures the standardization of case output through templates and rule constraints. At the same time, double verification ensures the accuracy of structured results, thus solving the core problem of the difficulty in balancing intelligence and standardization in existing technologies.

[0008] Furthermore, between the large model intelligent interpretation step and the configurable structural constraint step, a hierarchical dynamic verification embedding step is also included: the large model intelligent interpretation step is decomposed into entity extraction sub-steps, relation association sub-steps, and field mapping sub-steps; a basic syntax verification sub-step is set for the entity extraction sub-step to verify the legality of the extracted medical entity types and the completeness of key medical entities; a clinical logic verification sub-step is set for the relation association sub-step to verify the rationality of the extracted clinical relations; a format constraint verification sub-step is set for the field mapping sub-step to verify the format compliance and field length compliance of the medical entities mapped to the case structured template fields; when an anomaly is detected in any verification sub-step, a corresponding correction instruction is pushed to the large model, and the large model completes the correction of the medical entity or clinical relation in the current sub-step according to the correction instruction before proceeding to the next sub-step.

[0009] By adopting the above technical solution, the verification process is embedded into each sub-step of the large model interpretation, realizing the synchronous interpretation and verification. This avoids secondary corrections caused by post-verification, significantly improves the efficiency of case structure processing, and ensures that the processing results of each step meet the requirements, reducing the error rate of the final output.

[0010] Furthermore, in the configurable structural constraint step, the process of defining the case structure template includes: decomposing the field constraint rules into basic constraint cells, medical standard cells, clinical logic cells, and department-specific cells. The basic constraint cells are used to limit the basic field requirements of the case structure template; the medical standard cells are used to limit the association requirements between medical terminology and medical standards; the clinical logic cells are used to limit the logical requirements of clinical relationships; and the department-specific cells are used to limit the specific field requirements of cases in a specific department. The process also involves receiving department type information and case scenario information input by the user, filtering matching cells from various types of cells based on the department type information and case scenario information, and automatically assembling them to form a case structure template adapted to the department type and case scenario.

[0011] By adopting the above technical solution, the template rules are decomposed into functional cells and automatically assembled, eliminating the need for manual configuration of complete templates. This significantly improves the configuration efficiency of case templates, while meeting the personalized needs of different departments and scenarios, and solving the problem of poor adaptability of existing templates.

[0012] Furthermore, the large-scale model intelligent interpretation step also includes a case complexity adaptive adjustment step: constructing a case complexity assessment model, which takes the number of symptoms, disease rarity, number of complications, and historical error rate of similar cases in the preprocessed text as input features, and outputs a case complexity score; adjusting the interpretation parameters of the large-scale model according to the case complexity score, the interpretation parameters including feature extraction dimension and semantic matching threshold; if the case complexity score is in a first preset interval, then the first feature extraction dimension and the first semantic matching threshold are used; if the case complexity score is in a second preset interval, then the second feature extraction dimension and the second semantic matching threshold are used, where the second feature extraction dimension is greater than the first feature extraction dimension and the second semantic matching threshold is higher than the first semantic matching threshold; if the case complexity score is in a third preset interval, then the third feature extraction dimension and the third semantic matching threshold are used, where the third feature extraction dimension is greater than the second feature extraction dimension and the third semantic matching threshold is higher than the second semantic matching threshold, and initiating a clinical guideline matching sub-step, calling the clinical diagnosis and treatment guideline knowledge graph to verify the diagnosis and treatment related information output by the large-scale model.

[0013] By adopting the above technical solution, the large model interpretation parameters are dynamically adjusted according to the complexity of the case. A higher precision interpretation strategy is used for complex cases and guideline verification is added, while a lightweight strategy is used for simple cases. This ensures the accuracy of handling complex cases while avoiding the waste of resources for simple cases, thus improving the overall processing efficiency of the system.

[0014] Furthermore, the configurable structural constraint step also includes a temporal cell association step: constructing a mapping map of temporal nodes and functional cells based on the clinical diagnosis and treatment path. The temporal nodes include consultation completion nodes, examination request submission nodes, examination execution completion nodes, report generation and upload nodes, diagnosis confirmation nodes, and treatment plan formulation nodes. The completion status of each temporal node in the medical information system is monitored in real time. When any temporal node is completed, the activation or update of the corresponding functional cell is automatically triggered according to the mapping map. The update of the functional cell includes adjusting the functional cell constraint rules based on the temporal node association data and supplementing the corresponding field content of the case structure template.

[0015] By adopting the above technical solutions, the case templates are dynamically updated along with the clinical diagnosis and treatment process, ensuring that the structured cases are synchronized with the progress of diagnosis and treatment, avoiding the information lag problem caused by static templates, and improving the real-time nature and completeness of the structured cases.

[0016] Furthermore, in the large model intelligent interpretation step, a hierarchical reinforcement learning strategy is adopted to optimize the interpretation process of the large model. Specifically, this includes: decomposing the interpretation task of the large model into entity extraction sub-tasks, relation association sub-tasks, and field mapping sub-tasks; setting a corresponding reward mechanism for each sub-task, wherein the reward mechanism determines the reward value based on the completion quality of the sub-task; determining the total reward function based on the reward value of each sub-task; and using a reinforcement learning algorithm to update the interpretation strategy of the large model according to the total reward function, wherein the reinforcement learning algorithm maximizes the total reward value by adjusting the strategy parameters.

[0017] By adopting the above technical solution, hierarchical reinforcement learning is introduced into the interpretation process of large models. The interpretation strategy is optimized by sub-task rewards and total reward functions, which solves the problem of slow convergence caused by sparse rewards in traditional large model interpretation and improves the accuracy and convergence efficiency of large models in extracting medical information.

[0018] Furthermore, the configurable structural constraint step also includes a cross-hospital privacy-aware federated collaboration step: multiple hospital nodes initialize their local functional cell libraries and case structured template rules respectively; each hospital node determines its privacy budget and probability of participating in federated aggregation based on the privacy sensitivity of its local medical data, wherein the privacy sensitivity is negatively correlated with the privacy budget and the probability of participating in federated aggregation; each hospital node updates its local parameters using a dynamic step size, wherein the dynamic step size is determined based on the node's data volume and computing resource utilization; the server receives the local parameters from the hospital nodes participating in federated aggregation, performs privacy protection processing on the local parameters, aggregates them according to preset weights to form global parameters, and distributes the global parameters to each hospital node to achieve synchronization of the functional cell libraries and case structured template rules of each hospital node.

[0019] By adopting the above technical solution and combining privacy-aware strategies in the cross-hospital template collaboration process, it not only achieves the synchronous sharing of case templates between different hospitals, but also protects medical data privacy through privacy budget and dynamic step size, thus resolving the contradiction between privacy protection and template synchronization in cross-hospital collaboration.

[0020] Furthermore, the data acquisition and preprocessing steps also include a multimodal anomaly handling step: synchronously acquiring doctor-patient dialogue text, examination report images, and electronic medical record fragments to form multimodal data; using a robust filtering algorithm to process the time-series numerical data in the multimodal data, wherein the robust filtering algorithm adapts to non-Gaussian noise by dynamically adjusting the filtering parameters; and using an online weight learning algorithm to adjust the fusion weights of each modality in real time, wherein the fusion weights are determined based on the current error rate and feature consistency of each modality.

[0021] By adopting the above technical solutions, introducing multimodal data and performing anomaly protection, the integrity and quality of data input are improved. Robust filtering and online weight learning ensure the stability of data processing under non-Gaussian noise and modal quality fluctuations, providing high-quality input for subsequent large model interpretation.

[0022] Furthermore, the verification and output steps also include a causal interpretability verification step: using a causal representation learning algorithm to separate causal features and related features from medical data, wherein the causal features are features that have a direct causal impact on the structured case results; calculating the intervention effect of the causal features, and selecting causal features with intervention effects greater than a preset threshold as verification criteria; generating an attention visualization map of the causal features, wherein the attention visualization map is used to display the contribution weight of each causal feature to the structured case results.

[0023] By adopting the above technical solution, key causal features are separated and visualized through causal interpretability verification, which solves the "black box" problem of large model interpretation, makes the generation logic of structured results traceable, improves the trust of clinical medical staff in structured cases, and meets the compliance requirements of medical scenarios.

[0024] Secondly, this application provides a large-model-driven intelligent medical case full-process structured system. It includes: a medical data acquisition terminal for acquiring doctor-patient dialogue voice, examination report images, and electronic medical record fragments; the medical data acquisition terminal includes a voice acquisition device and an image scanning device; a clinical configuration terminal for receiving user-input department type information, case scenario information, and field constraint rule configuration instructions, and displaying a preview of the case structured template and verification results; a main server communicatively connected to the medical data acquisition terminal and the clinical configuration terminal; the main server includes a processor and a memory; the memory stores a computer program; when the processor executes the computer program, it implements all the steps of the large-model-driven intelligent medical case full-process structured method according to any one of claims 1-9; and a medical information system interface module for transmitting the verified structured cases to the hospital's medical information system in a preset data format.

[0025] By adopting the above technical solution, a dedicated hardware carrier is provided for the intelligent medical case structure method driven by large models. Each terminal and module has a clear division of labor and works in concert to ensure that the method can be stably implemented and achieve seamless integration of structured cases with the hospital's existing medical information system, thus meeting the needs of actual clinical applications.

[0026] In summary, this application has at least the following beneficial effects:

[0027] 1. A large-scale model-driven intelligent medical case structured method and system is provided to achieve efficient transformation of unstructured doctor-patient dialogues into standardized cases;

[0028] 2. Improve processing efficiency through hierarchical dynamic verification, and adapt to multi-department scenarios through automatic assembly of rule cells;

[0029] 3. Enhance clinical trust through causal interpretability verification, and balance sharing and privacy protection through cross-hospital privacy federation collaboration.

[0030] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0031] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0032] Figure 1 This illustration shows a schematic diagram of a large-model-driven intelligent medical case full-process structured system according to an embodiment of this application.

[0033] Figure 2 This paper illustrates a flowchart of a large-model-driven intelligent medical case structure method in an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0036] This application provides a large-model-driven intelligent medical case structured method and system, which can efficiently convert unstructured doctor-patient dialogues into standardized cases, taking into account both intelligent interpretation by large models and case standardization, adapting to multi-departmental scenarios, protecting privacy in cross-hospital collaboration, improving clinical trust, and meeting the needs of medical data processing.

[0037] In one aspect, embodiments of this application disclose a large-model-driven intelligent medical case full-process structured system.

[0038] Figure 1 The diagram illustrates the principle of a large-model-driven intelligent medical case full-process structured system according to an embodiment of this application.

[0039] Reference Figure 1 The system is a macro-level hardware operating environment built to realize the aforementioned large-model-driven intelligent medical case full-process structured method. It includes multiple functionally coordinated hardware modules. The whole system ensures that unstructured doctor-patient dialogues can be efficiently transformed into standardized structured cases through data interaction and logical linkage, and can be smoothly connected to the hospital's existing medical information system.

[0040] The core hardware components and collaborative relationships of the system are as follows: The medical data acquisition terminal, as the raw data input end, consists of a voice acquisition device (such as a microphone for outpatient consultation) and an image scanning device (such as a high-definition scanner for examination reports). Its main function is to collect raw medical data such as doctor-patient dialogue voice, examination report images, and electronic medical record fragments. The terminal establishes a connection with the main server through a preset communication protocol and transmits the collected raw data to the main server in real time, providing data source support for subsequent data preprocessing and large model interpretation.

[0041] The clinical configuration terminal, serving as a human-computer interaction interface, consists of a terminal device equipped with a monitor and input devices (mouse, keyboard). Clinical personnel can use this terminal to input department type information (such as pediatrics, cardiology), case scenario information (outpatient, emergency), and field constraint rule configuration commands. Simultaneously, the terminal can receive and display the case structured template preview information and verification result information fed back from the main server, realizing the issuance of configuration commands and the visual viewing of processing results. It maintains bidirectional communication with the main server to ensure real-time interaction between configuration requirements and feedback data.

[0042] The main server, as the core computing and control hub of the system, consists of server equipment equipped with a high-performance processor and a large-capacity memory. The memory pre-stores computer programs that implement the aforementioned large-model-driven intelligent medical case full-process structured method. By executing the program, the processor completes all core steps, including data preprocessing (redundancy filtering, terminology normalization), intelligent interpretation of the large model (medical entity extraction, clinical correlation analysis), configurable structural constraints (template generation, rule injection), and verification output (rule verification, model verification). The main server establishes communication connections with the medical data acquisition terminal, clinical configuration terminal, and medical information system interface module, and is the key node for realizing data flow and functional coordination among the modules.

[0043] The medical information system interface module, as a data output interface component, consists of a data transmission interface adapted to the hospital's medical information system (HIS / LIS system). Its function is to convert and transmit the verified structured medical records output by the main server according to the preset data format (such as JSON or XML) supported by the hospital information system, ensuring that the structured medical records can be directly integrated into the hospital's existing information management system. This module communicates with the main server on one end to receive structured medical record data, and interfaces with the hospital's medical information system on the other end to complete the data entry.

[0044] Each hardware module forms a complete operational loop through standardized communication links. From the collection and input of raw medical data to the interaction of configuration commands, the processing of core algorithms, and finally the output and docking of structured cases, the hardware implementation of a large model-driven intelligent medical case full-process structured method has been realized, providing stable and efficient macro-equipment support for the case structured needs in clinical scenarios.

[0045] Secondly, embodiments of this application disclose a large-model-driven intelligent medical case full-process structured method.

[0046] Figure 2 A flowchart of a large-model-driven intelligent medical case structure method is shown in an embodiment of this application.

[0047] Reference Figure 2 The method specifically includes the following steps:

[0048] S1: Data Acquisition and Preprocessing Steps: Acquire doctor-patient dialogue voice, convert the doctor-patient dialogue voice into text using medical customized speech recognition technology, and perform redundant information filtering, medical terminology normalization, and data format standardization on the text.

[0049] The medical-customized speech recognition technology optimizes the recognition model by constructing a medical-specific dictionary (covering ICD-11 disease terms, national essential drug list drug names, etc.), reducing recognition errors for professional terms such as "myocardial infarction" and "nifedipine," and ensuring the accuracy of speech-to-text conversion. In the redundant information filtering stage, the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm is used to quantify the importance of words in the text, filtering out redundant greetings (such as "Hello" and "Excuse me") and repetitive expressions (such as a patient's repeated phrase "It's been hurting for 3 days, it's been hurting constantly"). Specifically, this is achieved by calculating the TF-IDF value of the words. Implementation, in which Words In the text Frequency of occurrence in ( , For words In the text (number of times it appears in) Words Inverse document frequency ( This represents the total number of texts. For words containing (the number of texts), only retaining Words exceeding a preset threshold (usually set to 0.1) are filtered for redundant information. Medical terminology normalization is achieved by constructing a mapping table of "non-standard terms - standard terms" (e.g., "myocardial infarction"). The text "coronary atherosclerotic heart disease (acute myocardial infarction)" will be mapped to standardized terms in ICD-11 or the National Medical Terminology Standard Database; data format standardization will unify date formats (e.g., "24-5-10"). "2024-05-10" and numerical formats (e.g., "body temperature 38 degrees") "body temperature Ensure that the data format meets the access requirements of the hospital information system (HIS / LIS).

[0050] The method in this step also includes a multimodal anomaly handling step: synchronously collecting doctor-patient dialogue text, examination report images, and electronic medical record fragments to form multimodal data; using a robust filtering algorithm to process the time-series numerical data in the multimodal data, wherein the robust filtering algorithm adapts to non-Gaussian noise by dynamically adjusting the filtering parameters; and using an online weight learning algorithm to adjust the fusion weights of each modality in real time, wherein the fusion weights are determined based on the current error rate and feature consistency of each modality.

[0051] The robust filtering algorithm is specifically the Robust Kalman Filter (RKF), which replaces the Gaussian noise assumption of the traditional Kalman Filter with a Student t-distribution to adapt to the non-Gaussian jump noise commonly found in medical time-series numerical data (such as body temperature and blood routine indicators). Its likelihood function is:

[0052]

[0053] in, for Time series numerical observations at time (e.g.) (Patient's body temperature measurement at any time) for State estimates at any given time (such as filtered and corrected body temperature values). for The observation noise covariance at time point, The degrees of freedom of the student's t-distribution are updated in real time through maximum likelihood estimation. The smaller the value, the stronger the non-Gaussian nature of the noise and the stronger the resistance to anomalies. For gamma function, This is the observation matrix (usually set to 1 in medical time-series data, representing the direct observation state). To further reduce the impact of non-Gaussian outliers, RKF introduces robust weights. The observations are downweighted. The calculation formula is:

[0054]

[0055] Corresponding filter gain Adjusted to:

[0056]

[0057] in for The prior state covariance at time t is adjusted to ensure that the impact of outlier observations on the filtering results is significantly reduced.

[0058] The online weight learning algorithm aims to minimize the multimodal data fusion error. It constructs an objective function and updates the weights of each modality in real time using gradient descent. Assume the multimodal data includes doctor-patient dialogue text modality. Inspection report image OCR modality Electronic medical record fragment modality The feature vectors of each mode are respectively ( Indicates the first (Data samples), the manually labeled true feature vectors are Then the objective function is:

[0059]

[0060] in, For the number of data samples, For the first The fusion weights of each modality (satisfying) ), This is the regularization coefficient (empirically set to 0.01 to avoid fusion imbalance caused by excessive weights of a single modality). It is the L2 norm. To achieve online weight updates, the stochastic gradient descent (SGD) algorithm is used. After iterative optimization, the updated formula is as follows:

[0061]

[0062] in, For the number of iterations, Set the learning rate to 0.001 to balance update speed and stability. For the first After the nth iteration Each modal weight. During the weight update process... The value of is determined by the current error rate and feature consistency of each modality: if the current error rate of a certain modality (such as the OCR modality) is 1, then the value of is determined by the current error rate of each modality. Increase, or consistency with other high-reliability modes ( If the cosine similarity decreases, then the mode's It will automatically decrease through gradient descent to ensure that multimodal fusion is always dominated by high-quality modes.

[0063] S2: Large Model Intelligent Interpretation Steps: A large model finely tuned with parallel corpora in the medical field is used to perform semantic understanding on the preprocessed text, extracting medical entities and clinical relationships between entities. The medical entities include patient basic information entities, symptom entities, disease entities, drug entities, and examination entities. The clinical relationships include symptom-disease relationships, drug-disease relationships, and examination-disease relationships.

[0064] The method in this step also includes a case complexity adaptive adjustment step: constructing a case complexity assessment model, which takes the number of symptoms, disease rarity, number of complications, and historical error rate of similar cases in the preprocessed text as input features, and outputs a case complexity score; adjusting the interpretation parameters of the large model according to the case complexity score, the interpretation parameters including feature extraction dimension and semantic matching threshold; if the case complexity score is in a first preset interval, then the first feature extraction dimension and the first semantic matching threshold are used; if the case complexity score is in a second preset interval, then the second feature extraction dimension and the second semantic matching threshold are used, where the second feature extraction dimension is greater than the first feature extraction dimension and the second semantic matching threshold is higher than the first semantic matching threshold; if the case complexity score is in a third preset interval, then the third feature extraction dimension and the third semantic matching threshold are used, where the third feature extraction dimension is greater than the second feature extraction dimension and the third semantic matching threshold is higher than the second semantic matching threshold, and initiating a clinical guideline matching sub-step, calling the clinical diagnosis and treatment guideline knowledge graph to verify the diagnosis and treatment related information output by the large model. Meanwhile, a hierarchical reinforcement learning strategy is adopted to optimize the interpretation process of the large model, specifically including: decomposing the interpretation task of the large model into entity extraction sub-tasks, relation association sub-tasks, and field mapping sub-tasks; setting a corresponding reward mechanism for each sub-task, wherein the reward mechanism determines the reward value based on the completion quality of the sub-task; determining the total reward function based on the reward value of each sub-task; and using a reinforcement learning algorithm to update the interpretation strategy of the large model according to the total reward function, wherein the reinforcement learning algorithm maximizes the total reward value by adjusting the strategy parameters.

[0065] The parallel corpus in the medical field contains over 100,000 samples of "doctor-patient dialogue text - structured case annotation" covering outpatient, emergency, and inpatient scenarios. The annotation content covers medical entity boundaries (such as "patient basic information entity" including sub-entities such as name, age, and gender) and clinical relationship types (such as "symptom-disease association" annotated as "cough"). (Pneumonia). The large-scale model fine-tuning uses a joint loss function of "entity recognition + relation classification," the formula of which is:

[0066]

[0067] in, The loss weights are used to balance the importance of entity recognition and relationship classification tasks. Cross-entropy loss for entity recognition (based on the BIO annotation system, where B represents the beginning of an entity, 1 represents the interior of an entity, and O represents a non-entity, calculating the probability difference between the model's predicted label and the true label). Cross-entropy loss is used for relation classification (calculating the probability loss of the model's prediction of relation types such as "symptom-disease" and "drug-disease"). The main model architecture adopts a medical-customized BERT-BiLSTM-CRF model. The BiLSTM layer captures textual context dependencies, and the CRF layer optimizes entity boundary prediction to ensure that medical entities extract F1 scores. Accuracy of clinical correlation identification .

[0068] The method in this step also includes a case complexity adaptive adjustment step: constructing a case complexity assessment model, which takes the number of symptoms, disease rarity, number of complications, and historical error rate of similar cases in the preprocessed text as input features, and outputs a case complexity score; adjusting the interpretation parameters of the large model according to the case complexity score, the interpretation parameters including feature extraction dimension and semantic matching threshold; if the case complexity score is in a first preset interval, then the first feature extraction dimension and the first semantic matching threshold are used; if the case complexity score is in a second preset interval, then the second feature extraction dimension and the second semantic matching threshold are used, where the second feature extraction dimension is greater than the first feature extraction dimension and the second semantic matching threshold is higher than the first semantic matching threshold; if the case complexity score is in a third preset interval, then the third feature extraction dimension and the third semantic matching threshold are used, where the third feature extraction dimension is greater than the second feature extraction dimension and the third semantic matching threshold is higher than the second semantic matching threshold, and initiating a clinical guideline matching sub-step, calling the clinical diagnosis and treatment guideline knowledge graph to verify the diagnosis and treatment related information output by the large model.

[0069] The case complexity assessment model employs an XGBoost regression model, with all input features normalized to [value missing]. Interval, complexity score ( The formula for calculating points is:

[0070]

[0071] in, (Symptom number weights are determined using the AHP (Analytical Hierarchy Process) method, and are assigned based on the degree of clinical impact on case complexity.) (Disease rarity weighting, rare diseases are assigned values ​​according to the ICD-11 rare disease catalog.) Common disease assignment ), (Complications number weighted, no complications) Types and above complications ), (The historical error rate weights for similar cases are taken from the error statistics of similar diseases in historical structured data.) This is a correction term used to offset feature normalization bias. The preset interval is set based on clinical case classification criteria: First preset interval (Simple cases, such as the common cold), corresponding to the first feature extraction dimension of 384 dimensions and the first semantic matching threshold of 0.55; the second preset interval (Moderate cases, such as bacterial pneumonia), corresponding to a second feature extraction dimension of 768 dimensions and a second semantic matching threshold of 0.65; the third preset interval (For complex cases, such as acute myocardial infarction complicated with heart failure), the corresponding third feature extraction dimension is 1024 and the third semantic matching threshold is 0.75.

[0072] The clinical practice guideline knowledge graph is constructed based on the "Clinical Practice Guidelines (2024 Edition)" and includes a "disease-treatment-drug recommendation" triple (e.g., "acute myocardial infarction-percutaneous coronary intervention-aspirin"). During validation, the matching degree between the large model's output treatment information and the guideline knowledge is calculated.

[0073]

[0074] in, The first recommended by the guide Items of diagnosis and treatment (such as drug dosage and frequency of medication). The first output of the large model Items of diagnosis and treatment (Measure the degree of content overlap) Weighting of guideline items (drug dosage) Frequency of medication Treatment process ,like If this occurs, a guideline deviation warning will be triggered, and a fragment of the original guideline will be pushed to the large model for correction.

[0075] Meanwhile, a hierarchical reinforcement learning strategy is adopted to optimize the interpretation process of the large model, specifically including: decomposing the interpretation task of the large model into entity extraction sub-tasks, relation association sub-tasks, and field mapping sub-tasks; setting a corresponding reward mechanism for each sub-task, wherein the reward mechanism determines the reward value based on the completion quality of the sub-task; determining the total reward function based on the reward value of each sub-task; and using a reinforcement learning algorithm to update the interpretation strategy of the large model according to the total reward function, wherein the reinforcement learning algorithm maximizes the total reward value by adjusting the strategy parameters.

[0076] The subtask reward value is calculated as follows: Entity extracts subtask reward. ,in The set of medical entities predicted by the model. The set of real entities is manually labeled, and the F1 score represents the overall entity precision. ) and recall rate ( ); Rewards for related sub-tasks ,in This is the set of clinical associations predicted by the model. For the set of true relationships, accuracy measures the percentage of correctly predicted relationships; field mapping subtask rewards. ,in The content of the model is mapped to the case template field. The IOU value represents the actual field content, and it measures the content coverage.

[0077] The total reward function is weighted based on the importance of subtasks:

[0078]

[0079] Among them, the relation association subtask has the highest weight (0.4), as the rationality of clinical association directly affects the diagnostic and treatment value of cases; entity extraction (0.3) and field mapping (0.3) have balanced weights to ensure the integrity and format compliance of basic information. The reinforcement learning algorithm adopts PPO (Proximal Policy Optimization), which updates the objective function through a pruning and stabilization strategy.

[0080]

[0081] in, For the current policy network (output in state) Next action probability, action Examples include "adjusting entity extraction threshold" and "correcting relationship classification weights". This is the old policy network (used to compute the policy update step size constraint). The pruning factor (limits the policy update range and avoids training oscillations). For the advantage function ( For the value of the action, State value measures the advantage of the current action relative to the average action. Through this optimization, the number of samples required for large model interpretation convergence was reduced from over 100,000 to over 30,000, and the error rate in intermediate steps (such as missed entity detection) was reduced from... Down to .

[0082] This step also includes a hierarchical dynamic verification embedding step: the large model intelligent interpretation step is broken down into entity extraction sub-steps, relationship association sub-steps, and field mapping sub-steps; a basic syntax verification sub-step is set for the entity extraction sub-step to verify the legality of the extracted medical entity types and the completeness of key medical entities; a clinical logic verification sub-step is set for the relationship association sub-step to verify the rationality of the extracted clinical relationships; a format constraint verification sub-step is set for the field mapping sub-step to verify the format compliance and field length compliance of the medical entities mapped to the case structured template fields; when an anomaly is detected in any verification sub-step, a corresponding correction instruction is pushed to the large model, and the large model completes the correction of the medical entity or clinical relationship in the current sub-step according to the correction instruction before proceeding to the next sub-step.

[0083] In the basic syntax validation sub-step, the legality verification of medical entity types is achieved through a pre-trained medical entity classifier. This classifier outputs the probability that the entity to be verified belongs to each legal type (such as "symptom", "disease", "drug"). Legality is determined by the following formula:

[0084]

[0085] in, For medical entities to be verified (such as "cough"). For a predefined set of legal entity types ({patient basic information, symptoms, disease, medication, examination}), For classifiers to classify entities Belongs to type The characteristic output value, Transform the eigenvalues ​​into a probability distribution; if ( Based on a legality threshold determined by statistical analysis of over 100,000 historical entity annotation data, the entity type is deemed legal; otherwise, it is considered abnormal. The integrity verification of critical medical entities is achieved by calculating the coverage of critical entities, using the following formula:

[0086]

[0087] in, The set of key entities defined for the case template (e.g., outpatient cases should include "symptoms, disease, and course of disease"). The entity set extracted from the large model. Indicates the number of elements in the set; if If the core diagnostic and treatment information is not missing, then the key entity is determined to be missing, triggering an integrity anomaly prompt.

[0088] The clinical logic verification sub-step relies on a medical knowledge graph (containing over 100,000 standard associations between "symptoms" and "drugs" and derived from ICD-11, the National Essential Medicines List, and clinical guidelines). It comprehensively judges the reasonableness of the association by calculating the similarity between the predicted association and the standard association, as well as the prior probability of the standard association.

[0089]

[0090] in, Clinical associations extracted from large models (such as "cough") hypertension"), for The vector representation of (generated through a knowledge graph embedding model). For knowledge graphs and The most similar standard association (e.g., "cough-pneumonia"). for The vector representation of , Cosine similarity (measures the degree of similarity between vectors). For standard association Prior probability ( This reflects the frequency of this association in clinical practice; if ( If the lowest score for a reasonable association, as determined by clinical experts, is used, then the association is deemed unreasonable.

[0091] In the format constraint validation sub-step, field format compliance is achieved through regular expression matching, and the format compliance rate is calculated to quantify the validation result.

[0092]

[0093] in, This is a collection of all fields in the case template (such as "age", "diagnosis date", and "drug dosage"). For fields Regular expression matching functions for formatting (e.g., "age" must match) The age and diagnosis date must match. "), True indicates that the format is compliant; if If the majority of fields are formatted correctly, then a formatting error is detected. Field length compliance is determined by calculating the length deviation rate.

[0094]

[0095] in, Mapping large models to fields Content, Len For content length, For fields Maximum allowed length (e.g., "chief complaint") The characters are derived from the medical record writing standards. ;like If a small length deviation is allowed, then the field length is considered abnormal.

[0096] When any verification sub-step detects an anomaly, the generation logic of the correction instruction corresponds one-to-one with the anomaly type: when the entity type is an anomaly, push "Entity"... The type is invalid; it is recommended to correct it to a valid type. When a key entity is missing, push the "key entity" message. Missing information, please supplement and extract it; if the clinical correlation is unreasonable, push "Correlation 'r' is unreasonable, please refer to standard correlation ' "Correct"; "Push field if format / length is abnormal" 'Format / length error, please conform to the rules' After receiving the correction instruction, the large model calls the "anomaly correction module" learned during fine-tuning in the current sub-step, reprocesses the anomaly based on the compliance requirements indicated by the instruction, and then proceeds to the next sub-step after the correction is completed, thus preventing the anomaly from being passed on to subsequent processes.

[0097] S3: Configurable structural constraint steps: Define a case structure template using a visual configuration tool. The case structure template includes a field hierarchy structure and field constraint rules. The field constraint rules include mandatory field rules, field format rules, and medical standard association rules. Convert the field constraint rules into prompt information and inject it into the large model to constrain the large model to output structured cases.

[0098] The visual configuration tool uses a tree-structured node interface. Parent field nodes (such as "Patient Information" and "Treatment Information") and child field nodes (such as "Name," "Age," and "Diagnosis Result") are linked by dragging and dropping to define their hierarchical relationship. The tool includes a built-in "template preview window" to display the field combination effect in real time. In the field constraint rules, the required field rules quantify the constraint strength through "accumulated required field weights," using the following formula: ,in This is a set of required fields (such as "symptoms", "diagnosis", and "course of disease" for outpatient cases). For fields Clinical importance weighting of "diagnosis" ,"symptom" "Progress of the disease" ), For indicator functions (when the field value is empty) ,otherwise ),like (Based on the maximum allowable missing weight set according to clinical case quality control standards), the template configuration will prompt "Required field missing". Field formatting rules are verified through "format matching rate", the formula is: ,in A function to match field values ​​with preset regular expressions (such as matching "age"). Matching "age" and "date" "), It conforms to the format requirements. Medical standard association rules ensure compliance through "term mapping rate," the formula of which is... ,in A set of medical terms associated with the field. This is a standard terminology database for ICD-11, the National Essential Medicines List, etc. The association requirements are met. When converting constraint rules into prompts, a "rule-prompt" mapping logic is used. For example, the prompt for "Required field missing" is "Please fill in the following required fields: \{f\inF_\{ntext\{req}}\midI(f_\{text\{empty}})=1}". The prompts are injected into the large model context in natural language format, forcing the output content to conform to the template constraints.

[0099] In this step, the process of defining the case structure template includes: decomposing the field constraint rules into basic constraint cells, medical standard cells, clinical logic cells, and department-specific cells. The basic constraint cells are used to limit the basic field requirements of the case structure template; the medical standard cells are used to limit the association requirements between medical terminology and medical standards; the clinical logic cells are used to limit the logical requirements of clinical relationships; and the department-specific cells are used to limit the specific field requirements for cases in a specific department. The method also involves receiving department type information and case scenario information input by the user, filtering matching cells from various types of cells based on the department type information and case scenario information, and automatically assembling them to form a case structure template adapted to the department type and case scenario.

[0100] Cell selection is achieved through the calculation of "scene-cell matching degree", the formula is as follows: ,in For "department-scene" feature vectors (such as pediatric outpatient scene) (0.8 indicates pediatrics, 0.8 indicates outpatient). For feature vectors of functional cells (such as the "feeding history" of department-specific cells) ), Cosine similarity (measures the consistency of vector directions). Adapting weights to the cell scene (basic constraint cells) Department-specific cells ), The cell is identified as a matching cell. Automatic assembly is performed based on the "basic constraint cell". Medical Standard Cell Clinical logic cells The "department-specific cells" are executed in a priority order, ensuring that the template first meets basic requirements such as "field existence," and then advanced constraints such as "terminology compliance" and "logical rationality" are added. After assembly, it is verified through "template integrity score," the formula of which is... ,in Basic constraint cell coverage ( ), The coverage rates of the other three types of cells are respectively. Time templates can be used for structured output.

[0101] Simultaneously, it also includes a time-sequential cell association step: constructing a mapping map of time-sequential nodes and functional cells based on the clinical diagnosis and treatment pathway. The time-sequential nodes include consultation completion nodes, examination request submission nodes, examination execution completion nodes, report generation and upload nodes, diagnosis confirmation nodes, and treatment plan formulation nodes. The completion status of each time-sequential node in the medical information system is monitored in real time. When any time-sequential node is completed, the activation or update of the corresponding functional cell is automatically triggered according to the mapping map. The update of the functional cell includes adjusting the functional cell constraint rules based on the time-sequential node association data and supplementing the corresponding field content of the case structure template.

[0102] The completion status of time-series nodes is monitored in real time through a "status bit matrix". (6 corresponds to 6 timing nodes). Indicates the first When each node completes (e.g., when the "report generation and upload node" completes) The status bits are synchronized from the medical information system every 10 seconds. The mapping map uses a directed graph. Build, It is a set of nodes (including temporal nodes and functional cell nodes). For a set of edges (such as "report generation upload node") The edge of the "inspection result cell" represents the node's completion triggering cell activation, and the edge weight is... Indicates the trigger priority ("diagnostic confirmation node") "ICD-coded cell" This is higher than the "treatment plan formulation milestone". "Drug Cells" When updating a functional cell, the formula for adjusting the rule parameters is as follows: ,in This is the threshold for the original cell rule (such as the "threshold for judging abnormal results"). This is the average value of the data associated with the time-series nodes (such as the average value of a certain indicator in the uploaded report). To adjust the coefficients and ensure the thresholds adapt to real-time data distribution, field content supplementation is measured by the "data fill rate," calculated using the following formula: When this is not the case, it is considered as supplementary qualification (such as filling the key values ​​in the report into the "Inspection Results" field).

[0103] In addition, the process includes a cross-hospital privacy-aware federated collaboration step: multiple hospital nodes initialize their local functional cell libraries and case structured template rules respectively; each hospital node determines its privacy budget and probability of participating in federated aggregation based on the privacy sensitivity of its local medical data, wherein the privacy sensitivity is negatively correlated with the privacy budget and the probability of participating in federated aggregation; each hospital node updates its local parameters using a dynamic step size, wherein the dynamic step size is determined based on the node's data volume and computing resource utilization; the server receives the local parameters from the hospital nodes participating in federated aggregation, performs privacy protection processing on the local parameters, aggregates them according to preset weights to form global parameters, and distributes the global parameters to each hospital node to achieve synchronization of the functional cell libraries and case structured template rules of each hospital node.

[0104] Privacy sensitivity (No. (Each hospital node) is quantified through data type and privacy level, using the formula: ,in For nodes The dataset, For data Weights (tumor data) Common cold data ), For data Privacy levels (1-5, with level 1 being the lowest and level 5 being the highest). (After normalization). Privacy Budget and Negative correlation, the formula is ,in For maximum privacy budget ( hour), hour Probability of participation in aggregation Highly sensitive nodes Reduce the risk of privacy leaks. Dynamic step size. Calculated as ,in The initial step size, For nodes The amount of data, Total data volume CPU utilization This represents the load regulation coefficient. Privacy protection processing employs Laplace noise injection, as shown in the formula... ,in For nodes Local parameters, Parameter sensitivity (setting the final aggregation parameter based on the maximum range of parameter variation) This enables cross-departmental synchronization of cell libraries and template rules.

[0105] S4: Verification and Output Steps: Perform rule verification and model verification on the structured cases output by the large model. The rule verification is used to verify the compliance of medical terminology and the correctness of field format in the structured cases. The model verification is used to verify the rationality of clinical relationships in the structured cases. If the verification passes, the structured cases are output to the medical information system. If the verification fails, the large model is triggered to reprocess the structured cases or the manual review process is initiated.

[0106] In rule verification, the compliance of medical terminology is quantified using the "standard terminology matching rate," calculated using the following formula: ,in This is a collection of all medical terms included in a structured case (such as "myocardial infarction" and "aspirin"). To integrate the terminology databases of authoritative standards such as ICD-11 and the National Essential Medicines List, Indicates the number of elements in the set; if If the minimum compliance rate is set based on medical data quality control requirements, the terminology is deemed compliant; otherwise, it is marked as an abnormal term and the specific discrepancy is recorded (e.g., "myocardial infarction" does not match the standard term "coronary atherosclerotic heart disease (acute myocardial infarction)"). Field format correctness is assessed using the "format error rate," calculated using the formula: ,in For the collection of all fields of a structured case, For fields The value of , This function matches field values ​​against preset regular expressions (e.g., "age" must match). age"), (allow The format is considered correct if the formatting error is within a certain range.

[0107] Model validation comprehensively assesses the reasonableness of the association relationship through "clinical association confidence score", the formula of which is: ,in Clinical associations output by the large model (e.g., "fever") Vector representation of "pneumonia" (generated through a knowledge graph embedding model). This represents the vector representation of the most similar standard associations in a medical knowledge graph. Cosine similarity (measures the consistency of vector directions). Clinical confidence level of the standard association (calculated based on the frequency of the association in authoritative guidelines, such as "fever") The word "pneumonia" appeared 120 times in the guidelines. );like If the association reaches the minimum confidence level determined by clinical experts, then the association verification is successful. The comprehensive judgment rule for the verification result is: when... and and If the overall verification passes, the structured case is output to the hospital's HIS / LIS system in the HL7FHIR standard format (a common data exchange format for medical information systems). If any indicator fails to meet the standard, the large model is first reprocessed (calling the "error correction prompt" module, such as "Term 'myocardial infarction' is not compliant, please correct it to the standard term"). If it still fails to meet the standard after reprocessing (two consecutive failures), the manual review process is initiated, and the anomalies (such as incorrect format fields or low confidence associations) are marked and pushed to the clinician workstation.

[0108] The method in this step also includes a causal interpretability verification step: using a causal representation learning algorithm to separate causal features and related features from medical data, wherein the causal features are features that have a direct causal impact on the structured case results; calculating the intervention effect of the causal features, and selecting causal features with intervention effects greater than a preset threshold as verification criteria; generating an attention visualization map of the causal features, wherein the attention visualization map is used to display the contribution weight of each causal feature to the structured case results.

[0109] Causal representation learning algorithms employ causal graph models (CGM) and separate features through the following steps: First, a feature association graph is constructed. For feature nodes in the preprocessed text (such as "body temperature") "Cough" and "White blood cell count" ), The edges connecting features have weights. (Mutual information, measuring features) and (The correlation strength); then, based on do-calculus, causal features are identified, if the feature satisfy ( For structured case results, such as "Diagnosis: Pneumonia", then If it is a causal feature (such as "elevated white blood cell count" directly affecting the diagnosis of "pneumonia"), then it is a related feature (such as "patient gender" and "pneumonia" only having a statistical association, with no direct causal effect).

[0110] The intervention effect was quantified using the average treatment effect (ATE), as shown in the formula: ,in Indicates mandatory features Presence (e.g., "elevated white blood cell count is present"). Indicates mandatory features It does not exist. Let it be the expected function; When the causal characteristic is determined to have a significant impact on the outcome (based on an empirical threshold of clinical causal strength), it is included in the validation set.

[0111] The contribution weights of the attention visualization graph are calculated using feature importance, as shown in the formula: ,in For the set of all causal features, Representation of features Relative contribution weights (normalized to) The graph is displayed in the form of a heatmap, with features having higher weights appearing in darker colors (e.g., "elevated white blood cell count"). Displayed in dark red, "Cough" (Displayed in light red), this visually presents the degree of influence of each causal feature on the structured results, providing interpretable evidence for clinicians' review.

[0112] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0113] By applying medical-customized speech recognition technology in the data acquisition and preprocessing steps, the transcription error of professional medical terms can be reduced. Combined with redundant information filtering and medical terminology normalization, invalid information can be eliminated and expression standards can be unified, providing a high-quality text foundation for subsequent processing. Multimodal anti-anomaly processing, through robust filtering algorithms to adapt to non-Gaussian noise and online weight learning to dynamically adjust fusion weights, can improve the consistency and anti-interference ability of multi-source data, further ensuring the reliability of input data.

[0114] Based on preprocessed data, the large model, fine-tuned with parallel corpora in the medical field, can more accurately capture the domain characteristics of medical entities and clinical relationships during intelligent interpretation. Adaptive adjustment of case complexity optimizes feature extraction dimensions and semantic matching thresholds by evaluating the complexity score output by the model, enabling the large model to maintain appropriate interpretation accuracy when handling simple, medium, and complex cases. Hierarchical reinforcement learning, through sub-task reward mechanisms and overall reward function optimization strategies, can specifically improve the quality of sub-tasks such as entity extraction, relationship association, and field mapping. The combined effect of these three aspects can improve the completeness of medical entity recognition and the accuracy of clinical relationship extraction.

[0115] The hierarchical dynamic validation embedding step sets up corresponding validation steps after entity extraction, relationship association, and field mapping sub-steps. This can promptly intercept issues such as abnormal entity types, missing key entities, unreasonable association logic, and incorrect format lengths in the process. By pushing correction instructions in real time, the large model can be corrected at the current stage, effectively preventing errors from being passed on to subsequent steps and ensuring the compliance of intermediate results.

[0116] In the configurable structural constraint step, the visual configuration tool and cell disassembly and assembly mechanism can generate structured templates that adapt to departments and scenarios, ensuring that the output framework meets actual clinical needs; the temporal cell association is based on the dynamic activation or updating of functional cells according to the diagnosis and treatment path, which can make the template adapt to the diagnosis and treatment process in real time and improve the timeliness of structured content; the cross-hospital privacy-aware federated collaboration can achieve cross-hospital rule synchronization under the premise of protecting privacy by controlling participation, dynamically updating parameters by step size and federated aggregation, and enhancing the universality and standardization of templates.

[0117] The rule verification of the validation and output steps, through terminology compliance and format correctness verification, ensures that structured cases meet medical standards and system access requirements; model validation, through clinical association confidence assessment, ensures the clinical rationality of the association; causal interpretability verification, through separating causal features, calculating intervention effects, and generating attention maps, improves the interpretability and clinical trustworthiness of the results; the reprocessing and manual review mechanism for validation failures further guarantees the quality of the output.

[0118] In summary, the end-to-end technical approach, from data input to result output, systematically improves the accuracy, compliance, timeliness, and clinical applicability of structured case results by optimizing data quality, enhancing interpretation accuracy, intercepting intermediate errors, standardizing the output framework, and strengthening final verification. This enables a reliable transformation from doctor-patient dialogue to standardized cases. The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the aforementioned disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A large-model-driven intelligent medical case structured method, characterized in that, Includes the following steps: Data acquisition and preprocessing steps: Acquire doctor-patient dialogue voice, convert the doctor-patient dialogue voice into text through medical customized speech recognition technology, and perform redundant information filtering, medical terminology normalization and data format standardization on the text. The large-scale intelligent interpretation steps are as follows: A large-scale model finely tuned with parallel corpora in the medical field is used to perform semantic understanding on the preprocessed text, extracting medical entities and clinical relationships between entities. The medical entities include patient basic information entities, symptom entities, disease entities, drug entities, and examination entities. The clinical relationships include symptom-disease relationships, drug-disease relationships, and examination-disease relationships. Configurable structural constraint steps: Define a case structure template using a visual configuration tool. The case structure template includes a field hierarchy and field constraint rules. The field constraint rules include mandatory field rules, field format rules, and medical standard association rules. Convert the field constraint rules into prompt information and inject it into the large model to constrain the large model to output structured cases. Verification and output steps: The structured cases output by the large model are subjected to rule verification and model verification. The rule verification is used to verify the compliance of medical terminology and the correctness of field format in the structured cases. The model verification is used to verify the rationality of clinical relationships in the structured cases. If the verification passes, the structured cases are output to the medical information system. If the verification fails, the large model is triggered to reprocess the structured cases or the manual review process is initiated.

2. The large-model-driven intelligent medical case structured method according to claim 1, characterized in that, Between the large model intelligent interpretation step and the configurable structural constraint step, a hierarchical dynamic verification embedding step is also included: The intelligent interpretation steps of the large model are broken down into entity extraction sub-steps, relationship association sub-steps, and field mapping sub-steps. A basic syntax validation sub-step is set up for the entity extraction sub-step to verify the legality of the extracted medical entity types and the integrity of key medical entities; A clinical logic verification sub-step is set up for the corresponding relationship association sub-step to verify the rationality of the extracted clinical association relationship; A format constraint validation sub-step is set for the field mapping sub-step to verify the format compliance and field length compliance of the medical entity mapping to the medical record structure template fields; When an anomaly is detected in any verification sub-step, a corresponding correction instruction is pushed to the large model. The large model then corrects the medical entity or clinical relationship in the current sub-step according to the correction instruction before proceeding to the next sub-step.

3. The large-model-driven intelligent medical case structured method according to claim 2, characterized in that, The process of defining the case structure template in the configurable structural constraint step includes: The field constraint rules are broken down into basic constraint cells, medical standard cells, clinical logic cells, and department-specific cells. The basic constraint cells are used to limit the basic field requirements of the case structure template. The medical standard cells are used to limit the association requirements between medical terminology and medical standards. The clinical logic cells are used to limit the logical requirements of clinical associations. The department-specific cells are used to limit the specific field requirements of cases in a specific department. The system receives department type information and case scenario information input by the user. Based on the department type information and case scenario information, it selects matching cells from various types of cells and automatically assembles them to form a structured case template that is adapted to the department type and case scenario.

4. The large-model-driven intelligent medical case structured method according to any one of claims 1-3, characterized in that, The large-scale model intelligent interpretation step also includes a case complexity adaptive adjustment step: A case complexity assessment model is constructed, which takes the number of symptoms, disease rarity, number of complications, and historical error rate of similar cases in the preprocessed text as input features and outputs a case complexity score. The interpretation parameters of the large model are adjusted based on the case complexity score. The interpretation parameters include the feature extraction dimension and the semantic matching threshold. If the case complexity score is within a first preset range, then a first feature extraction dimension and a first semantic matching threshold are used; if the case complexity score is within a second preset range, then a second feature extraction dimension and a second semantic matching threshold are used, wherein the second feature extraction dimension is greater than the first feature extraction dimension and the second semantic matching threshold is higher than the first semantic matching threshold; if the case complexity score is within a third preset range, then a third feature extraction dimension and a third semantic matching threshold are used, wherein the third feature extraction dimension is greater than the second feature extraction dimension and the third semantic matching threshold is higher than the second semantic matching threshold, and a clinical guideline matching sub-step is initiated, calling the clinical diagnosis and treatment guideline knowledge graph to verify the diagnosis and treatment-related information output by the large model.

5. The large-model-driven intelligent medical case structured method according to any one of claims 1-4, characterized in that, The configurable structural constraint step also includes a temporalized cell association step: A mapping map of temporal nodes and functional cells is constructed based on the clinical diagnosis and treatment pathway. The temporal nodes include the consultation completion node, the examination application submission node, the examination execution completion node, the report generation and upload node, the diagnosis confirmation node, and the treatment plan formulation node. The system monitors the completion status of each time-series node in the medical information system in real time. When any time-series node is completed, the corresponding functional cell is automatically activated or updated according to the mapping map. The update of the functional cell includes adjusting the functional cell constraint rules based on the time-series node association data and supplementing the corresponding field content of the case structure template.

6. The large-model-driven intelligent medical case structured method according to any one of claims 1-5, characterized in that, In the large model intelligent interpretation step, a hierarchical reinforcement learning strategy is used to optimize the interpretation process of the large model, specifically including: The task of interpreting the large model is broken down into entity extraction subtask, relationship association subtask, and field mapping subtask. A corresponding reward mechanism is set for each subtask, and the reward value is determined based on the quality of completion of the subtask; The total reward function is determined based on the reward value of each sub-task. A reinforcement learning algorithm is then used to update the interpretation strategy of the large model according to the total reward function. The reinforcement learning algorithm maximizes the total reward value by adjusting the policy parameters.

7. The large-model-driven intelligent medical case structured method according to any one of claims 1-6, characterized in that, The configurable structural constraint step also includes a cross-departmental privacy-aware federated collaboration step: Multiple hospital nodes initialize their local functional cell libraries and case structure template rules respectively; Each hospital node determines its privacy budget and probability of participating in federated aggregation based on the privacy sensitivity of its local medical data. The privacy sensitivity is negatively correlated with the privacy budget and the probability of participating in federated aggregation. Each of the hospital nodes updates its local parameters using a dynamic step size, which is determined based on the node's data volume and computing resource utilization. The server receives local parameters from hospital nodes participating in the federated aggregation, performs privacy protection processing on the local parameters, aggregates them according to preset weights to form global parameters, and distributes the global parameters to each hospital node to achieve synchronization of the functional cell library and case structure template rules of each hospital node.

8. The large-model-driven intelligent medical case structured method according to any one of claims 1-7, characterized in that, The data acquisition and preprocessing steps also include a multimodal anomaly mitigation step: Simultaneously collect doctor-patient dialogue texts, examination report images, and electronic medical record fragments to form multimodal data; A robust filtering algorithm is used to process the time-series numerical data in the multimodal data. The robust filtering algorithm adapts to non-Gaussian noise by dynamically adjusting the filtering parameters. An online weight learning algorithm is used to adjust the fusion weights of each modality in real time. The fusion weights are determined based on the current error rate and feature consistency of each modality.

9. The large-model-driven intelligent medical case structured method according to any one of claims 1-8, characterized in that, The verification and output steps also include a causal interpretability verification step: A causal representation learning algorithm is used to separate causal features and correlation features from medical data. The causal features are those that have a direct causal impact on the structured results of cases. Calculate the intervention effect of the causal features, and select causal features with intervention effects greater than a preset threshold as verification criteria; An attention visualization map of causal features is generated, which is used to show the contribution weight of each causal feature to the structured case outcome.

10. A large-model-driven intelligent medical case structured system, characterized in that, include: A medical data acquisition terminal is used to collect voice conversations between doctors and patients, images of examination reports, and fragments of electronic medical records. The medical data acquisition terminal includes a voice acquisition device and an image scanning device. The clinical configuration terminal is used to receive user input of department type information, case scenario information, and field constraint rule configuration instructions, and to display the case structure template preview information and verification result information. A main server is communicatively connected to the medical data acquisition terminal and the clinical configuration terminal. The main server includes a processor and a memory. The memory stores a computer program. When the processor executes the computer program, it implements all the steps of the large model-driven intelligent medical case full-process structured method according to any one of claims 1-9. The medical information system interface module is used to transmit verified structured medical records to the hospital's medical information system in a preset data format.

Images